Generative and Agentic Artificial Intelligence in Banking

Abstract

Artificial intelligence is increasingly reshaping the organisation and delivery of financial services. While traditional AI has long been applied to areas such as fraud detection, risk assessment and algorithmic decision-making, the emergence of generative AI and increasingly agentic systems is extending AI from task-specific automation towards knowledge-work augmentation, workflow integration and potentially autonomous action. This paper examines how these developments may transform banking operations and what organisational and governance challenges they create, using Raiffeisen as a case study. Drawing on the Raiffeisen case and recent academic literature on generative AI, AI agents, financial services and human–AI collaboration, the paper analyses the opportunities and risks associated with increasing levels of AI integration.

The analysis finds that generative AI can create substantial value through improved information access, knowledge-work augmentation, productivity gains and process support. Agentic AI extends this potential by enabling AI systems to retrieve information, reason across multiple steps and potentially execute actions within defined workflows. However, increasing AI capability does not automatically translate into organisational value. The effectiveness of AI depends on data quality, organisational preparedness, employee AI literacy, workflow design, calibrated trust and robust governance. Moreover, greater autonomy introduces additional risks relating to reliability, data privacy, cybersecurity, bias, regulatory compliance and accountability.

The paper argues that the transformation of banking should therefore be understood not as a linear progression towards replacing human work, but as a reconfiguration of work and decision-making through human–AI collaboration. For institutions such as Raiffeisen, the strategic challenge is to match the level of AI autonomy to the risk and consequences of the task being performed. Sustainable competitive advantage is consequently likely to depend less on access to AI technology itself than on an institution's ability to integrate increasingly capable AI systems with proprietary data, organisational knowledge, human expertise and effective governance. The paper concludes that responsible AI adoption requires innovation and control to develop together: the value of increasing AI autonomy ultimately depends on the organisational capability to govern it.

Keywords: Generative AI; Agentic AI; Banking; Financial Services; Artificial Intelligence; Raiffeisen; Digital Transformation

1. Introduction

Artificial intelligence (AI) has evolved from a specialised analytical technology into a strategic component of digital transformation across the financial services industry. Banks have employed machine learning and related AI techniques for many years in areas such as fraud detection, credit assessment, risk management, customer analytics and algorithmic decision-making. Recent developments in generative artificial intelligence (GenAI), however, represent a potentially more consequential technological shift. Unlike conventional AI systems that are typically designed to identify patterns, make predictions or execute narrowly defined tasks, GenAI can generate new text, code, summaries, analyses and other forms of content in response to natural-language instructions. Its significance therefore extends beyond the automation of individual processes to the augmentation and potential reconfiguration of knowledge-intensive work (Vuković, Dekpo-Adza and Matović, 2025; Eisfeldt and Schubert, 2025).

This distinction is particularly important in banking because financial institutions are fundamentally information-processing organisations. A substantial proportion of banking work involves interpreting large volumes of structured and unstructured information, applying complex regulatory requirements, preparing documents, communicating with customers and colleagues, and transforming organisational knowledge into decisions and actions. A recent systematic review of AI applications in banking identifies operational efficiency, customer service, financial decision-making, explainability and regulatory adoption as major themes in the literature, demonstrating that AI is increasingly embedded across the functional architecture of financial institutions (Singh, 2025). GenAI extends this trajectory by bringing AI capabilities into activities that have traditionally depended heavily on human language, reasoning and professional expertise.

The emerging literature consequently suggests that the significance of GenAI should not be understood solely in terms of cost reduction or task automation. Eisfeldt and Schubert (2025) argue that financial occupations are particularly exposed to the productivity effects of generative AI because many activities within finance involve information processing and knowledge-intensive tasks. Similarly, Ganesh et al. (2025) conceptualise generative AI agents as a means of augmenting knowledge work in finance, emphasising the potential for AI systems to support skilled professionals in processing information, reasoning over complex problems and executing tasks. This perspective shifts the analytical focus from the question of whether AI will replace employees towards the more consequential question of how work itself will be reorganised around increasingly capable AI systems.

The emergence of agentic AI marks a further development in this transformation. Whereas a conventional GenAI application generally responds to a user instruction by producing an output, agentic systems are designed to pursue objectives through multiple steps, potentially combining reasoning, information retrieval, tool use and action. Ganesh et al. (2025) argue that such systems could progressively augment financial knowledge work while allowing the degree of AI autonomy to increase under appropriate forms of human–AI collaboration. The distinction is strategically significant for banks: moving from AI that generates information to AI that can act upon information potentially changes not only employee productivity but also the design, coordination and governance of organisational processes.

At the same time, the adoption of GenAI in banking cannot be considered a purely technological or productivity-oriented development. Financial institutions operate in an environment characterised by extensive regulation, sensitive customer data, high requirements for operational reliability and significant consequences arising from erroneous decisions. Empirical research on GenAI adoption in banking highlights managerial preparedness, reliability, regulatory compliance, data privacy and organisational responsiveness as central conditions for successful implementation (Moharrak and Mogaji, 2025). More broadly, systematic research on AI in financial services identifies regulatory challenges, explainability, governance and responsible adoption as persistent issues accompanying the technological opportunities (Vuković, Dekpo-Adza and Matović, 2025). The strategic value of GenAI therefore depends not simply on what the technology is capable of doing, but on how effectively its capabilities can be embedded within appropriate organisational, regulatory and human-control structures.

The case of Raiffeisen provides a timely empirical lens through which these developments can be examined. As described by the IFZ Retail Banking Blog, the Raiffeisen Group is pursuing the use of generative AI, Microsoft Copilot and emerging forms of agentic AI across its organisation. The case is particularly relevant because it illustrates the movement from experimentation with generative AI towards its integration into everyday knowledge work and, potentially, towards increasingly autonomous forms of process execution (Dietrich, 2026). Rather than representing an isolated technological initiative, the Raiffeisen case can therefore be situated within a broader transformation of banking in which AI is progressively moving from a specialised analytical capability towards an organisational infrastructure for information processing, employee augmentation and process automation.

Against this background, this paper examines how generative and agentic AI may reshape banking operations, using Raiffeisen as a case through which the opportunities and challenges of this transformation can be explored. It addresses the following research question:

How can generative AI and agentic AI transform banking operations, and what organisational, technological and regulatory challenges must financial institutions such as Raiffeisen address to implement these technologies responsibly?

2. Artificial Intelligence and the Transformation of Banking

Artificial intelligence has become embedded in the technological infrastructure of modern financial services, extending from specialised analytical applications to increasingly broad forms of organisational and knowledge work. Banks were among the early adopters of machine-learning techniques because their core activities involve the systematic processing of large volumes of financial, behavioural and transactional data. AI has consequently been applied to fraud detection, credit assessment, risk management, customer analytics, forecasting and operational decision-making. A recent systematic review of 157 studies on AI in banking confirms the breadth of this development, identifying applications across customer service, financial operations, adoption and inclusion, regulation and explainability (Singh et al., 2025). The transformation is therefore not confined to a particular banking function; rather, AI is becoming a general-purpose technology with implications for how financial institutions create, deliver and govern services.

The economic significance of this development extends beyond operational efficiency. Recent research suggests that AI innovation can affect the broader performance of banking institutions. Gyau et al. (2024), for example, find a positive relationship between AI innovation and banks’ financial performance, while also emphasising the importance of complementary investment in information and communication technologies and appropriate regulatory frameworks. Such findings indicate that the value of AI does not arise automatically from deploying algorithms. Instead, technological investment must be accompanied by organisational capabilities and institutional conditions that allow financial institutions to translate technological capabilities into measurable improvements.

The emergence of generative AI introduces a qualitatively different dimension to this transformation. Earlier generations of banking AI were predominantly designed to perform defined analytical functions: detecting anomalous transactions, predicting credit risk or classifying customers, for example. Generative AI, by contrast, can interact with users through natural language and generate new content, including text, summaries, code and analytical outputs. This substantially broadens the range of activities that can be supported by AI because many previously human-centred tasks involve language, interpretation and the synthesis of information rather than purely numerical prediction. Eisfeldt and Schubert (2025) argue that financial occupations are particularly exposed to the productivity effects of generative AI precisely because a significant proportion of financial work consists of information-intensive and knowledge-based activities.

This distinction is particularly important in banking. Financial institutions are not simply data-processing organisations; they are also knowledge-intensive organisations in which employees continuously interpret regulations, analyse customer and market information, prepare documentation, communicate with stakeholders and apply professional judgement. Generative AI can therefore intervene directly in the production and transformation of organisational knowledge. Rather than merely automating a discrete task, it can assist employees across a sequence of activities, from retrieving relevant information and summarising documents to drafting communications and supporting analytical work. The potential impact is consequently organisational as much as technological.

This perspective is consistent with the emerging literature on AI-enabled knowledge-work augmentation. Ganesh et al. (2025) argue that generative AI agents have the potential to augment financial knowledge workers by supporting cognitive activities that have traditionally depended on human expertise. The central implication is not necessarily the wholesale substitution of employees, but a redistribution of tasks between humans and AI systems. Employees may increasingly delegate routine information-processing activities to AI while retaining responsibility for interpretation, judgement, relationship management and consequential decisions. Such a shift could alter the composition of banking jobs even where employment itself is not directly displaced.

The implications extend further as generative AI becomes increasingly integrated into organisational workflows. The transition from standalone AI tools to embedded AI assistants changes the nature of interaction between employees, information and enterprise systems. Instead of searching manually across multiple sources, employees may increasingly interact with organisational knowledge through natural-language interfaces. This has the potential to reduce the friction associated with information retrieval and documentation while simultaneously increasing the importance of the quality, accessibility and governance of the underlying organisational data. AI adoption therefore creates a reciprocal relationship: the usefulness of generative systems depends partly on the quality of the information environment into which they are embedded, while their introduction may in turn change how that information is accessed and used.

The next stage of this development is represented by increasingly agentic forms of artificial intelligence. Whereas a conventional generative AI application primarily produces an output in response to a prompt, AI agents can potentially perform sequences of tasks towards a defined objective. Recent research on intelligent financial systems highlights the possibility that generative AI and AI agents could affect core functions of the financial system, including intermediation, insurance, asset management and payments (Aldasoro et al., 2025). In an organisational context, this suggests a movement from AI as an assistant that provides information towards AI as an active participant in workflows that can retrieve information, reason over it, interact with digital systems and initiate actions.

This progression changes the strategic question facing banks. The issue is no longer simply whether AI can perform an existing task more quickly or cheaply. The more consequential question is whether organisations can redesign processes around new forms of human–AI collaboration. Such redesign may generate productivity gains, but it may also alter established responsibilities, skills and control structures. As AI becomes more capable of undertaking multi-step activities, banks must determine which decisions can appropriately be delegated, which require human validation and how accountability can be maintained when AI-generated outputs influence consequential financial processes.

At the same time, the transformation remains constrained by the institutional characteristics of banking. Financial services operate under stringent requirements concerning data protection, security, reliability, transparency and regulatory compliance. Moharrak and Mogaji (2025), based on empirical interviews with banking managers and industry experts, identify recognition, requirement, reliability, regulatory considerations and responsiveness as critical factors influencing the successful integration of generative AI in banking. Their findings reinforce the proposition that technological capability alone does not determine adoption. The ability to integrate GenAI effectively depends on managerial preparedness, organisational requirements and the capacity to manage the risks associated with the technology.

Explainability and accountability are particularly important because the consequences of errors in financial services can extend beyond operational inconvenience to financial loss, regulatory breaches and reputational damage. Singh et al. (2025) identify explainability and regulation as central themes in the AI-in-banking literature, suggesting that the increasing sophistication of AI does not eliminate traditional governance challenges but instead makes them more significant. The introduction of generative and agentic systems therefore creates a tension between increasing autonomy and the need for human control. The greater the capacity of an AI system to act independently, the greater the importance of mechanisms that ensure its actions remain observable, contestable and subject to appropriate oversight.

AI transformation in banking should consequently be understood as an organisational transformation rather than simply a technological upgrade. Traditional AI has primarily expanded banks’ capacity to analyse data and automate defined decisions. Generative AI extends this capability into language-based and knowledge-intensive work, while agentic AI potentially extends it further into the execution and coordination of multi-step processes. The resulting transformation concerns not only efficiency, but also the distribution of work between humans and machines, the organisation of knowledge, the design of business processes and the governance of decision-making.

This broader perspective provides the foundation for analysing Raiffeisen’s approach to generative AI, Copilot and agentic AI. The case is particularly valuable because it illustrates how these technological developments are translated from general technological possibilities into organisational practice. Examining Raiffeisen therefore makes it possible to assess not only what GenAI and agentic AI can theoretically achieve in banking, but also how a major financial institution is attempting to integrate these technologies into its existing structures, workflows and responsibilities.

3. The Raiffeisen Case: From Generative AI to Agentic AI

The Raiffeisen case provides a particularly relevant illustration of how generative and increasingly agentic forms of artificial intelligence are being translated from technological developments into organisational practice. As documented by the IFZ Retail Banking Blog, Raiffeisen is pursuing several forms of AI adoption, including generative AI applications, Microsoft Copilot and the exploration of agentic AI. The significance of this development lies not simply in the adoption of individual AI tools, but in the broader trajectory from AI-assisted employee productivity towards increasingly integrated and potentially autonomous forms of knowledge work (Dietrich, 2026).

The Raiffeisen approach is consistent with the broader transformation of AI in financial services. As discussed in Chapter 2, the initial generation of AI applications in banking was predominantly concerned with specialised analytical and predictive tasks. Generative AI expands this scope by enabling employees to interact with AI through natural language and by producing outputs that can directly support knowledge-intensive activities. The Raiffeisen case is therefore significant because it illustrates the movement of AI into the everyday working environment of employees rather than restricting its use to narrowly defined analytical applications.

A central element of this development is the use of AI to support employees rather than treating AI primarily as a customer-facing technology. Generative AI can assist with activities such as retrieving and synthesising information, drafting documents, summarising material and supporting analytical work. This orientation towards employee augmentation is consistent with the emerging literature on generative AI and knowledge work. Ganesh et al. (2025) argue that generative AI agents are particularly relevant to finance because financial knowledge work involves substantial cognitive processing of information by skilled professionals. From this perspective, the value of AI is not necessarily derived from replacing human expertise, but from allowing employees to delegate selected cognitive tasks while retaining responsibility for judgement and decision-making.

This distinction is important for understanding the potential significance of Microsoft Copilot within an organisation such as Raiffeisen. An AI assistant embedded within the digital environment in which employees already work can reduce the friction associated with information retrieval, drafting, summarisation and other routine knowledge activities. The resulting productivity effect may therefore arise not from a single automated process, but from the cumulative reduction of relatively small amounts of time spent on information-intensive tasks. This is consistent with the broader observation by Eisfeldt and Schubert (2025) that financial occupations are particularly exposed to the productivity effects of generative AI because of their high degree of information and knowledge intensity.

However, the productivity potential of enterprise AI depends on more than the capabilities of the underlying language model. The usefulness of an AI assistant is closely connected to the quality, accessibility and governance of the information to which it has access. In a banking environment, this becomes particularly important because employees operate with confidential customer information, internal policies, regulatory documentation and other sensitive organisational data. The integration of AI into everyday workflows therefore creates a dual requirement: organisations must make relevant knowledge accessible enough for AI systems to generate useful outputs while maintaining sufficiently strong controls over what information can be accessed, processed and disclosed.

The Raiffeisen case consequently illustrates an important shift in the locus of AI adoption. Rather than deploying AI only as a discrete application, the organisation is exploring how AI can become embedded within the broader digital workplace. This distinction matters because embedded AI has the potential to influence how employees search for information, produce documents, communicate and perform routine analytical activities. AI thus becomes part of the infrastructure through which organisational knowledge is accessed and transformed. The long-term implications may therefore extend beyond productivity to the organisation of work itself.

The move towards agentic AI represents a further development of this trajectory. Conventional generative AI generally operates reactively: a user provides an instruction and the system generates an output. Agentic AI introduces a greater degree of goal-directed behaviour, in which a system can potentially decompose an objective into multiple steps, retrieve relevant information, interact with software tools and execute actions. Ganesh et al. (2025) describe this development in terms of progressively increasing levels of autonomy and emphasise that financial knowledge work may ultimately be supported by agents capable of undertaking increasingly complex sequences of activities.

This distinction is particularly consequential in banking because the transition from generating information to taking action changes the risk profile of AI. An erroneous summary may create inconvenience or lead an employee to reconsider an analysis; an autonomous system that incorrectly executes a transaction, modifies a record or provides an inappropriate customer response could produce substantially more serious consequences. Agentic AI therefore introduces a fundamental tension between autonomy and control. Increasing autonomy may generate greater efficiency, but it simultaneously increases the importance of monitoring, authorisation, auditability and human intervention.

Recent research on AI in finance reinforces the significance of this transition. Aldasoro et al. (2025) argue that GenAI and emerging AI agents have the potential to affect core functions of the financial system, including financial intermediation, insurance, asset management and payments. Their analysis places agentic AI within a much broader transformation of financial information processing, rather than viewing it merely as an extension of office productivity software. A recent scientific review similarly identifies applications of AI agents across financial domains including fraud detection, credit risk assessment, financial advice and regulatory compliance, while highlighting the associated technical, ethical and regulatory challenges (Rizinski and Trajanov, 2026).

The Raiffeisen case should nevertheless be interpreted with an appropriate degree of caution. The emergence of agentic AI does not mean that autonomous AI systems can immediately replace established banking processes or professional judgement. Rather, the case illustrates an organisation positioning itself for a gradual expansion of AI capabilities. This distinction is important because the current state of agentic AI remains characterised by limitations in reliability, reasoning, contextual understanding and controllability. Ganesh et al. (2025) consequently emphasise architectures for human–AI collaboration that permit autonomy to increase progressively rather than assuming that fully autonomous systems are immediately appropriate for complex financial work.

From this perspective, the strategic importance of Raiffeisen's AI activities lies in the progression between different forms of human–AI interaction. Generative AI can initially function as an assistant that produces and transforms information at the request of an employee. Enterprise assistants such as Copilot can embed these capabilities directly into established workflows. Agentic AI potentially represents the next step, in which AI systems can initiate and coordinate sequences of actions towards defined objectives. Each step increases the potential productivity benefit, but it also increases the importance of organisational governance and human oversight.

The Raiffeisen case therefore provides an instructive example of a broader transformation occurring across financial services. AI is moving from a technology that primarily analyses data towards one that can participate in knowledge production and, increasingly, in the execution of organisational processes. The strategic challenge for Raiffeisen is consequently not simply to identify tasks that AI can perform. It is to determine where AI-generated assistance creates genuine organisational value, where human judgement remains indispensable, and under what conditions greater levels of AI autonomy can be introduced without compromising reliability, accountability and regulatory control.

The case thus supports a broader interpretation of AI adoption in banking: the fundamental transformation is not from human work to machine work, but from human-only workflows towards increasingly sophisticated forms of human–AI collaboration. Raiffeisen's progression from generative AI and Copilot towards agentic AI can be understood within this wider transformation, in which the boundaries between information generation, decision support and process execution are becoming increasingly fluid.

4. The Potential of Agentic AI in Financial Services

Agentic artificial intelligence represents a significant development in the evolution of generative AI because it extends AI systems beyond the generation of information towards the execution and coordination of actions. While conventional generative AI typically responds to a user instruction by producing an output, agentic systems can potentially combine reasoning, information retrieval, tool use and sequential task execution in pursuit of a defined objective. This distinction is particularly consequential in financial services, where many activities consist not of isolated decisions but of interconnected processes involving data retrieval, analysis, judgement and action. The emergence of agentic AI therefore raises the possibility that AI could move from supporting individual tasks to participating directly in the execution of financial workflows (Ganesh et al., 2025; Aldasoro et al., 2025).

The potential significance of this development derives partly from the structure of financial work itself. Financial institutions operate through complex information flows in which employees must continuously collect, interpret and reconcile information from multiple sources. Regulations, contracts, customer records, market information and internal policies may all need to be considered in a single process. Ganesh et al. (2025) argue that generative AI agents are particularly well suited to knowledge work in finance because they can potentially support professionals across multiple stages of information processing. Rather than simply generating a response, an agent can potentially retrieve relevant information, synthesise it, compare alternatives and support the subsequent execution of a task. This creates the possibility of shifting AI's role from a productivity tool towards an active component of organisational workflows.

The distinction between assistance and agency is therefore central. A conventional AI assistant may help an employee identify relevant information or draft a document, while an agentic system could potentially take responsibility for a sequence of related activities. For example, rather than merely summarising a regulatory update, an agent could potentially identify the relevant internal policies, assess which processes may be affected, prepare an initial assessment and route the issue to the appropriate employee for review. Similarly, in operational processes, an agent could potentially monitor defined conditions, retrieve relevant data and initiate predefined actions without requiring a separate human instruction at every stage. The potential productivity gains consequently arise not only from accelerating individual tasks but from reducing the coordination costs between tasks.

This potential is particularly relevant to financial operations. Aldasoro et al. (2025) provide an illustrative example by examining generative AI agents in the context of liquidity management within payment systems. Their analysis demonstrates that AI agents can perform sophisticated decision-making tasks in a simulated financial environment, including balancing competing objectives associated with liquidity management. Although such evidence should not be interpreted as demonstrating that autonomous AI is ready to manage real-world financial systems without supervision, it provides evidence that agentic systems can perform more complex financial reasoning than the narrow task automation associated with earlier generations of AI.

The significance of these developments extends beyond individual productivity. Aldasoro et al. (2025) argue that AI may ultimately affect the organisation of financial intermediation itself, while Ganesh et al. (2025) highlight the potential for AI agents to augment professional knowledge work. Taken together, these perspectives suggest that the long-term implications of agentic AI may concern the architecture of financial processes rather than merely the efficiency of individual employees. If AI systems become capable of retrieving information, reasoning over it and executing actions across multiple systems, established workflows may increasingly be redesigned around human–AI collaboration.

Such transformation could produce substantial efficiency benefits. Routine and repetitive activities could increasingly be delegated to AI agents, allowing employees to concentrate on activities requiring professional judgement, interpersonal interaction and responsibility. In areas such as compliance, for example, an agent could potentially assist with monitoring regulatory developments, identifying potentially relevant requirements and preparing documentation for human review. In customer operations, agents could potentially support the coordination of information across systems and prepare responses based on approved organisational knowledge. In financial analysis, they could potentially assemble data, conduct preliminary analyses and present alternatives to professional decision-makers. The common feature across these applications is not the elimination of human expertise but the redistribution of cognitive and operational tasks between humans and machines.

However, the potential benefits of increased autonomy must be considered alongside a corresponding increase in risk. An AI system that merely generates a draft for human review creates a different risk profile from one that can independently access systems and execute actions. As the autonomy of an AI system increases, the consequences of erroneous reasoning, incomplete information or inappropriate action may also increase. In financial services, where decisions can affect customers, markets and regulatory compliance, this creates a fundamental tension between automation and control.

Reliability is therefore a central challenge. Generative AI systems can produce outputs that appear plausible while nevertheless containing factual or logical errors. In an agentic context, such errors may propagate through a sequence of actions rather than remaining confined to a single generated response. A flawed interpretation at the beginning of an automated workflow could consequently influence subsequent decisions and actions. Ganesh et al. (2025) therefore emphasise the importance of architectures that support controlled human–AI collaboration and allow the level of autonomy to be adapted to the nature and risk of the task.

The problem of accountability becomes equally important. When an AI system generates a recommendation, responsibility can relatively clearly remain with the employee who evaluates it. When an AI agent performs several actions autonomously, however, the allocation of responsibility becomes more complex. The relevant question is not simply whether an AI system made an error, but whether the organisation established appropriate parameters, monitoring mechanisms and approval requirements for the system. Consequently, agentic AI requires governance mechanisms that extend beyond conventional model validation towards continuous monitoring of AI behaviour and the actions performed by AI-enabled systems.

Data governance represents another critical consideration. Agentic systems may require access to multiple internal and external information sources in order to perform complex tasks. In banking, these sources can contain highly sensitive customer, financial and operational information. The expansion of AI access therefore increases the importance of controls governing data permissions, information provenance and system boundaries. An agent must not only produce an appropriate answer; it must also operate within clearly defined limits concerning which information it can access and which actions it is authorised to perform.

Regulatory requirements further reinforce the need for controlled autonomy. Financial institutions operate within highly regulated environments in which decisions must often be explainable, auditable and attributable to identifiable individuals or organisational processes. The broader literature on AI adoption in financial services identifies regulation, explainability, data governance and accountability as persistent challenges associated with the increasing use of AI (Vuković, Dekpo-Adza and Matović, 2025; Singh et al., 2025). Agentic AI intensifies these challenges because autonomous systems potentially create a greater distance between the original human instruction and the eventual action.

This does not imply that agentic AI should be excluded from high-value financial processes. Rather, it suggests that autonomy should be treated as a variable that must be calibrated to the characteristics and risk of each task. Low-risk and highly structured activities may be suitable for greater levels of automation, whereas decisions involving material financial consequences, regulatory interpretation or sensitive customer outcomes may require explicit human approval. Ganesh et al. (2025) similarly emphasise the importance of designing human–AI systems in which autonomy and human involvement can be appropriately balanced.

For a banking organisation such as Raiffeisen, this creates a strategic challenge that extends beyond selecting an AI technology. The critical issue is to determine where agentic capabilities create genuine organisational value and where human expertise and accountability must remain central. This requires consideration of the task being performed, the quality and sensitivity of the information involved, the consequences of potential errors and the degree of autonomy that can reasonably be permitted. In this sense, the question is not whether an AI agent can perform a particular activity, but whether it should be permitted to perform that activity autonomously within the institution's risk and governance framework.

The potential of agentic AI in financial services should therefore be understood as both an opportunity and a governance challenge. Its distinctive contribution lies in the possibility of connecting information retrieval, reasoning and action within a single system, thereby enabling AI to participate in increasingly complex workflows. Yet precisely because agentic AI can move beyond generating information towards taking action, its adoption requires stronger mechanisms of oversight than conventional generative AI. The strategic value of agentic AI will ultimately depend on whether financial institutions can capture its productivity and process-innovation benefits while preserving reliability, accountability, regulatory compliance and human control.

For Raiffeisen, this balance is likely to determine whether the transition from generative AI and Copilot towards agentic AI becomes a meaningful organisational transformation or remains primarily an experimental technological initiative. The challenge is therefore not simply to increase the autonomy of AI systems, but to develop forms of autonomy that are proportionate to risk, transparent in operation and embedded within clearly defined human responsibilities.

5. Risks and Challenges of Generative AI in Banking

The transformative potential of generative artificial intelligence in banking is accompanied by a distinctive set of technological, organisational, regulatory and systemic risks. These risks do not necessarily arise because AI is inherently less reliable than other technologies; rather, they reflect the particular characteristics of generative systems. Unlike conventional software, large language models can generate plausible outputs without guaranteeing their factual accuracy, operate through probabilistic rather than deterministic processes, and may be difficult for users to interpret or challenge. In a highly regulated industry in which decisions can have significant financial and social consequences, these characteristics create challenges that extend beyond conventional information-technology risk management.

One of the most immediate concerns is the reliability of AI-generated outputs. Generative AI systems can produce information that is linguistically convincing but factually inaccurate, a phenomenon commonly described as hallucination or confabulation. This presents a particularly serious problem in financial services because an incorrect output may be incorporated into customer communication, financial analysis, regulatory interpretation or internal decision-making. Eisfeldt and Schubert (2025) identify the high exposure of financial occupations to generative AI, but this same exposure increases the importance of ensuring that AI-generated outputs are sufficiently reliable for their intended use. The central problem is therefore not simply whether an AI model can generate useful information, but whether users can reliably distinguish useful outputs from plausible but incorrect ones.

The empirical evidence reinforces this concern. Research evaluating the capabilities of generative AI in financial decision-making identifies important limitations in the ability of these systems to provide reliable financial guidance, particularly in contexts where users may lack sufficient financial or digital literacy (Khan et al., 2025). The risk is amplified when users attribute greater competence or authority to an AI system than is warranted. Crisanto et al. (2024) identify this phenomenon as an important feature of generative AI risk in financial services, noting that anthropomorphic characteristics can encourage users to treat AI systems as more knowledgeable or trustworthy than they actually are. The resulting problem is therefore partly technological and partly behavioural: even a system with known limitations can create significant risk if employees or customers fail to account for those limitations.

This issue becomes more consequential when generative AI is integrated into decision-making processes rather than used simply for drafting or information retrieval. If an employee treats an AI-generated analysis as authoritative without independently verifying the underlying information, human oversight may become nominal rather than substantive. The concept of human-in-the-loop governance therefore cannot be reduced to requiring a person to approve an AI output. Effective oversight requires that the human decision-maker possesses sufficient information, expertise and time to critically evaluate what the system has produced. Otherwise, the presence of a human reviewer may provide the appearance of control without materially reducing the underlying risk.

A second major challenge concerns data privacy and confidentiality. Banks operate with some of the most sensitive categories of organisational and customer information, including account data, transaction histories, identity information, financial circumstances and internal risk assessments. The use of generative AI creates questions about where this information is processed, how it is retained, what third-party systems can access it and whether information provided to an AI system can subsequently influence outputs in ways that compromise confidentiality. Moharrak and Mogaji (2025) identify data privacy as one of the central considerations affecting the adoption of generative AI in banking, alongside reliability and regulatory compliance.

The data challenge is particularly significant for enterprise AI because greater usefulness often depends on broader access to organisational information. An AI assistant that cannot access relevant internal information may provide limited value, whereas an AI system with extensive access may introduce significant privacy and security risks. Banks therefore face a fundamental governance challenge: AI systems must have sufficient access to perform useful tasks while remaining subject to strict controls concerning data minimisation, permissions and information boundaries. The development of agentic AI intensifies this issue because autonomous systems may require access to multiple data sources and software environments in order to complete multi-step tasks.

Cybersecurity represents a related but distinct challenge. Generative AI can strengthen financial institutions' defensive capabilities by assisting with threat detection, security analysis and the identification of anomalous activity. At the same time, the technology can lower the cost of producing sophisticated phishing messages, social-engineering content and malicious code. Aldasoro et al. (2024) highlight this dual character in their analysis of generative AI and cybersecurity in central banking: GenAI can support cyber defence while simultaneously creating new vulnerabilities associated with data disclosure and malicious use. The technology should therefore be understood as both a security capability and a potential source of additional attack vectors.

The risks associated with AI can also extend beyond individual errors to the stability of financial institutions and, potentially, the financial system. This is particularly important because widespread adoption may create correlated dependencies on similar models, data providers or technology platforms. Zhao, Dai and Nagayasu (2025), examining the introduction of ChatGPT in the Chinese banking sector, find evidence of an immediate increase in systemic financial risk following its emergence, with transitional challenges associated with GenAI adoption identified as an important mechanism. Their findings do not establish that GenAI necessarily destabilises banking systems in the long term, but they demonstrate that the process of technological transition itself can create financial vulnerabilities.

This systemic dimension complicates the traditional view of AI as an internal efficiency technology. If many financial institutions rely on similar foundation models, cloud providers, data sources or AI-enabled decision processes, a common technological failure could potentially affect multiple institutions simultaneously. Furthermore, widespread use of similar models may create correlated errors or behavioural responses. The resulting risk is qualitatively different from an isolated model failure because the consequences may propagate across institutions and markets.

A further challenge concerns bias and the quality of decisions supported by generative AI. Large language models learn patterns from extensive training data, which may contain historical biases, incomplete representations or problematic associations. In financial contexts, these biases can become consequential when AI systems are used to support investment advice, customer segmentation, lending decisions or other activities that affect individuals economically. Winder, Hildebrand and Hartmann (2025), for example, find that large language models can reinforce existing investment biases and increase portfolio risks among private investors. This demonstrates that the risks of generative AI are not limited to factual errors; AI-generated outputs can also systematically influence human behaviour in undesirable directions.

Regulatory compliance consequently represents another fundamental challenge. Banks already operate within extensive frameworks governing consumer protection, data privacy, operational resilience, model risk and financial conduct. The introduction of GenAI does not necessarily make these existing risks obsolete; rather, it changes how they may arise and how they need to be controlled. Crisanto et al. (2024) argue that existing financial regulation addresses many of the risks associated with AI, but also identify areas requiring further attention, including governance, expertise and skills, model risk management, data governance and dependence on third-party AI providers. This suggests that the central regulatory challenge is not simply the creation of entirely new rules, but the adaptation of established governance principles to systems whose behaviour can be less predictable and more difficult to validate than traditional financial models.

The question of accountability becomes particularly important in this context. Traditional financial models generally operate within relatively well-defined development, validation and monitoring processes. Generative AI systems can be more dynamic and difficult to evaluate because their outputs are qualitative, context-dependent and potentially variable. When an AI-generated recommendation contributes to a consequential decision, responsibility must therefore remain attributable to identifiable individuals and organisational processes. The introduction of agentic AI makes this issue even more important because AI systems may eventually execute actions rather than merely provide recommendations. As autonomy increases, governance must address not only the validity of model outputs but also the authorisation, monitoring and auditability of AI-generated actions.

Recent research on algorithmic governance in banking reinforces this point. García-Llorente and Olmeda (2026) argue that the increasing embedding of AI into banking processes creates challenges concerning model risk, governance responsibilities, documentation and enforceable accountability. Their analysis suggests that broad principles such as “trustworthy AI” are insufficient unless translated into concrete supervisory and organisational practices. This is particularly relevant to agentic systems because accountability cannot be established retrospectively if an institution cannot reconstruct what an AI system did, which information it used or why a particular action was taken.

These challenges indicate that generative AI cannot be treated as an ordinary software deployment. Traditional IT governance focuses substantially on system availability, cybersecurity, access control and technical reliability. GenAI requires these controls but also demands additional mechanisms addressing output quality, model behaviour, prompt and data governance, human reliance, validation and ongoing monitoring. The governance challenge therefore shifts from controlling a fixed technological system towards managing a probabilistic and evolving socio-technical system in which humans and AI jointly produce outcomes.

For a bank such as Raiffeisen, this distinction is particularly important. The successful deployment of GenAI depends not only on selecting appropriate models and applications but also on establishing clear boundaries around their use. Employees need to understand what AI systems can and cannot reliably do, which information may be entered into them, when outputs require independent verification and which decisions cannot appropriately be delegated. Governance must therefore operate at multiple levels: technological controls must be complemented by organisational policies, employee training, risk assessment and clearly assigned accountability.

The challenge becomes even more pronounced as banks move from generative AI towards agentic AI. A system that drafts a document can be reviewed before it is used; a system that independently executes a sequence of actions requires controls throughout the process. Agentic AI therefore increases the importance of permission structures, action limits, monitoring, audit trails and mechanisms for human intervention. The principle of proportionality becomes central: the greater the potential impact of an AI-enabled action, the stronger the requirements for human oversight and organisational control should be.

The risks of generative and agentic AI should therefore not be interpreted as arguments against adoption. Rather, they demonstrate that the value of AI in banking is inseparable from the quality of its governance. The literature increasingly points towards the need to integrate technological innovation with risk management, organisational capability and regulatory oversight (Moharrak and Mogaji, 2025; Vuković, Dekpo-Adza and Matović, 2025). The strategic challenge is to create conditions in which banks can exploit the productivity and innovation potential of AI without allowing technological capability to outpace institutional control.

For Raiffeisen, this means that the transition towards increasingly capable AI systems must be accompanied by a corresponding development of governance capabilities. Reliability, data protection, cybersecurity, bias, regulatory compliance and accountability should not be treated as secondary constraints added after implementation. They are fundamental conditions for determining where and how AI can create sustainable value. Ultimately, responsible AI adoption in banking requires the institution to govern not only the technology itself, but also the relationship between AI systems, employees, customers, data and organisational decision-making.

6. Human-AI Collaboration and Organisational Change

The organisational implications of generative and agentic AI extend beyond the automation of individual tasks. A central question is how the division of labour between employees and AI systems will evolve as AI becomes increasingly capable of performing cognitive and knowledge-intensive activities. The evidence emerging from recent research suggests that the most consequential transformation may not be the wholesale replacement of human workers, but the reconfiguration of work around new forms of human–AI collaboration. In this perspective, AI becomes an organisational capability that changes what employees do, how they perform it and how work is coordinated, rather than simply a substitute for human labour.

The distinction between automation and augmentation is therefore fundamental. Automation seeks to transfer the performance of a task from a human to a machine, whereas augmentation involves combining the respective strengths of humans and AI. Generative AI is particularly suited to the latter because it can perform information-intensive activities while leaving humans responsible for interpretation, judgement and accountability. Ganesh et al. (2025) argue that generative AI agents have significant potential to augment knowledge work in finance, where professionals routinely process complex information and apply domain-specific expertise. From this perspective, the economic value of AI depends not only on the tasks it can perform independently, but also on how effectively it can complement human capabilities.

Empirical evidence increasingly supports this augmentation perspective. Brynjolfsson, Li and Raymond (2025), analysing data from 5,172 customer-support agents, find that access to a generative AI assistant increased worker productivity by approximately 15% on average. Importantly, the effects were heterogeneous: less experienced and lower-skilled workers experienced particularly strong improvements in both speed and quality, while the most experienced workers experienced smaller gains in speed and some decline in quality. The findings suggest that AI does not affect all employees in the same way and that its organisational consequences depend on the interaction between the technology, the employee and the task being performed.

This heterogeneity is particularly relevant to banking. Financial institutions employ professionals with widely varying levels of expertise, from operational employees and customer-service staff to compliance specialists, relationship managers, analysts and senior decision-makers. AI may therefore have different effects across these occupational groups. For less experienced employees, AI could provide access to organisational knowledge and best-practice guidance that would otherwise require years of experience to develop. For highly experienced professionals, by contrast, AI may offer less additional value on routine tasks while creating new requirements for reviewing and challenging AI-generated outputs. The introduction of AI can consequently alter not only productivity but also the relative importance of different forms of human expertise.

The concept of the “jagged technological frontier” provides a further explanation for this variation. Dell'Acqua et al. (2025) demonstrate that AI can improve human performance substantially on some knowledge-intensive tasks while reducing performance on others. In their field experiment involving 758 knowledge workers, participants using GPT-4 completed tasks within the AI system's capability frontier faster and with higher quality, but performed worse on a complex managerial task outside that frontier. This finding has important implications for banking because it challenges the assumption that AI assistance produces uniformly positive outcomes. Employees need to understand not only how to use AI, but also when AI is likely to be reliable and when independent human reasoning should take precedence.

Consequently, effective human–AI collaboration requires more than simply providing employees with access to an AI system. It requires employees to develop the ability to assess AI outputs critically, identify uncertainty and recognise the boundaries of the technology's competence. AI literacy therefore becomes an important organisational capability. Recent research on human–GenAI teams finds that AI knowledge and perceived value influence trust in AI collaboration, while perceptions of accuracy play a particularly important role in determining whether employees trust AI systems (2025). For financial institutions, this suggests that employee training should extend beyond technical instructions on how to operate an AI tool. Employees must also understand the limitations, risks and appropriate applications of AI.

Trust is particularly important because effective collaboration requires neither complete distrust nor unconditional reliance. Excessive distrust may prevent employees from exploiting useful AI capabilities, while excessive trust can result in automation bias, in which employees accept AI-generated recommendations without sufficient scrutiny. Recent research on trust in human–AI collaboration in finance emphasises that trust has implications extending beyond technology adoption to fiduciary responsibility, reputation and potentially systemic financial stability (Mirabile, Corazza and Alonso-Moral, 2026). The challenge is therefore to establish calibrated trust: employees should trust AI to the extent justified by the system's demonstrated capabilities, the task context and the available evidence.

The importance of calibrated trust is especially pronounced in banking because AI-generated outputs can affect decisions involving customers' financial circumstances. An employee may use AI to summarise a customer file, prepare a communication or identify relevant regulatory requirements, but the responsibility for a consequential decision cannot simply be transferred to the AI system. Human oversight must therefore remain meaningful rather than merely procedural. The employee needs both the authority and the competence to challenge an AI-generated recommendation. This creates a direct connection between AI adoption and organisational capability: institutions cannot expect employees to exercise effective oversight if they have not been trained to understand the technology they are supervising.

Human–AI collaboration can also affect employee motivation and the nature of work itself. Wu et al. (2025), in four experimental studies involving 3,562 participants, find that collaboration with GenAI can improve task performance while simultaneously reducing intrinsic motivation. This finding is significant because productivity is only one dimension of organisational performance. If AI removes large portions of employees' problem-solving or creative responsibilities, employees may experience their work as less meaningful even where measurable efficiency increases. Conversely, if AI removes routine work and allows employees to focus on more complex and socially meaningful activities, it could potentially improve job quality. The organisational consequences therefore depend partly on how work is redesigned around AI rather than simply on how much work AI can automate.

This suggests that successful AI implementation requires deliberate job and workflow redesign. Organisations need to determine which activities should be automated, which should be augmented and which should remain primarily human. In banking, this distinction may be particularly important because activities differ substantially in their levels of complexity, regulatory sensitivity and interpersonal significance. Routine document processing may be highly suitable for AI assistance, whereas relationship management, complex credit decisions or sensitive customer interactions may continue to require substantial human involvement. The introduction of AI should therefore be accompanied by a reassessment of roles and responsibilities rather than simply an additional technological layer being placed on existing processes.

Organisational learning is another critical component of this transformation. As AI capabilities evolve rapidly, the skills required to work effectively with AI systems are unlikely to remain static. Employees need opportunities to learn through experimentation, feedback and experience, while organisations need mechanisms for identifying successful use cases and disseminating them across the institution. The evidence from Brynjolfsson, Li and Raymond (2025) is relevant in this respect because their study finds evidence that AI assistance can facilitate worker learning, particularly among less experienced employees. AI can therefore function not only as a productivity technology but also as a mechanism through which organisational knowledge and expertise are more rapidly distributed.

For a banking organisation such as Raiffeisen, this creates an important strategic opportunity. If generative AI can make organisational knowledge more accessible to employees, it could reduce some of the traditional barriers associated with acquiring expertise. A less experienced employee might use an AI assistant to navigate internal procedures, identify relevant information or obtain guidance on routine cases, while more experienced employees could focus on exceptions and higher-value decisions. However, this benefit depends on the quality of the knowledge embedded in the AI system and on employees understanding that AI-generated guidance remains subject to verification. The organisation must therefore ensure that AI complements institutional knowledge rather than becoming an uncontrolled substitute for it.

Human–AI collaboration also has implications for organisational coordination. AI can potentially function not only as an individual productivity tool but as a coordination technology connecting information, people and processes. Recent organisational research argues that generative AI can potentially contribute to collective intelligence when designed to support coordination and knowledge sharing rather than merely individual productivity (2025). This is particularly relevant to banking organisations, where processes frequently cross departmental boundaries. AI could potentially reduce information fragmentation by helping employees locate relevant knowledge and coordinate activities across functions.

Nevertheless, increased AI collaboration may also create new organisational dependencies. If employees become heavily reliant on AI systems for information retrieval, drafting or decision support, important forms of professional expertise may gradually weaken. This creates a potential paradox: AI can increase short-term productivity while simultaneously reducing opportunities for employees to develop the underlying skills that AI performs on their behalf. The problem is particularly relevant for junior employees, who traditionally develop expertise by repeatedly performing relatively routine tasks before progressing to more complex responsibilities. If those foundational tasks are extensively automated, organisations may need to develop alternative mechanisms through which employees acquire professional competence.

The transition towards agentic AI intensifies this organisational question. Generative AI assistants primarily support employees in performing tasks, whereas agentic systems may increasingly perform sequences of tasks with limited intervention. As discussed in Chapters 3 and 4, this could generate significant efficiency gains but would also require a clearer allocation of responsibilities between human employees and AI agents. Organisations may increasingly need employees whose role is not to perform every individual task themselves, but to configure, supervise, evaluate and intervene in AI-enabled processes. This would represent a significant change in the nature of managerial and professional work.

Such a transition also reinforces the importance of governance as an organisational capability. Human oversight cannot operate independently of institutional structures. Employees need clear policies defining when AI may be used, what information may be provided to AI systems, when outputs require verification and which decisions require explicit human approval. Managers need mechanisms for monitoring the performance of AI-enabled workflows, while risk and compliance functions need sufficient technical understanding to evaluate the associated risks. AI governance is therefore not simply a technical function; it becomes an organisational capability distributed across multiple roles.

The Raiffeisen case illustrates the importance of this broader perspective. The introduction of generative AI and Copilot should not be interpreted solely as a programme for improving employee productivity. It potentially represents a change in how employees interact with information, how knowledge is distributed and how organisational processes are executed. The longer-term introduction of agentic AI could extend this transformation by changing the boundary between human decision-making and automated execution. The success of such a transformation will therefore depend on Raiffeisen's ability to develop complementary capabilities in AI literacy, workflow redesign, employee development, governance and human oversight.

Human–AI collaboration should consequently be understood as a socio-technical transformation rather than a simple technology adoption exercise. The empirical evidence suggests that AI can generate meaningful productivity gains, but those gains are heterogeneous and depend strongly on the relationship between the technology, the task and the employee (Brynjolfsson, Li and Raymond, 2025; Dell'Acqua et al., 2025). At the same time, effective collaboration requires calibrated trust, appropriate AI literacy and organisational structures that preserve meaningful human judgement (Mirabile, Corazza and Alonso-Moral, 2026). For financial institutions, the central challenge is therefore to design an environment in which AI strengthens rather than erodes human expertise.

The implications for Raiffeisen are consequently broader than employee training alone. Successful AI transformation requires a deliberate redesign of work around complementary human and machine capabilities. Employees must be equipped to use AI effectively, challenge it when necessary and remain accountable for decisions within their professional remit. Managers must redesign workflows and performance expectations, while the organisation as a whole must ensure that increasing technological capability is matched by corresponding governance and learning capabilities. The long-term competitive advantage of AI in banking may therefore depend less on access to the technology itself—which is increasingly widespread—and more on an institution's ability to organise effective, trustworthy and productive collaboration between humans and increasingly capable AI systems.

7. Discussion

The Raiffeisen case illustrates a broader transformation taking place across the banking sector: artificial intelligence is moving from specialised analytical applications towards increasingly integrated forms of generative and potentially agentic AI. This development represents more than an incremental improvement in banking technology. It potentially changes the relationship between employees, organisational knowledge and business processes. The progression from generative AI assistance through embedded workplace applications towards increasingly autonomous AI agents suggests that the strategic significance of AI lies not only in its ability to perform individual tasks, but in its potential to reshape how work is organised and executed.

The analysis presented in this paper indicates that generative AI can create value in banking through several complementary mechanisms. At the employee level, it can reduce the time required for information retrieval, summarisation, drafting and other knowledge-intensive activities. At the organisational level, AI can make dispersed knowledge more accessible and facilitate the coordination of information across processes. At the process level, increasingly agentic systems may eventually be capable of performing sequences of activities with limited human intervention. The potential therefore extends from productivity enhancement to the redesign of organisational workflows.

The empirical literature provides support for this interpretation, but also demonstrates that the benefits of AI are neither automatic nor uniform. Brynjolfsson, Li and Raymond (2025), for example, find substantial productivity gains from generative AI assistance among customer-support workers, but also significant differences according to worker experience. Similarly, Dell'Acqua et al. (2025) demonstrate that AI can improve the speed and quality of knowledge work within its capability frontier while reducing performance on tasks that lie outside that frontier. These findings challenge a technologically deterministic view in which increasing AI capability necessarily produces uniformly positive organisational outcomes. Instead, the impact of AI depends on the interaction between the technology, the task, the employee and the organisational context.

This observation is particularly important for interpreting the Raiffeisen case. The introduction of generative AI and Copilot should not be evaluated simply according to whether these technologies can perform particular activities. Their organisational value depends on whether they improve the overall process in which those activities are embedded. An AI system that produces an excellent summary may create little value if employees still need to spend substantial time verifying its content or transferring the result between disconnected systems. Conversely, relatively modest AI capabilities may generate substantial value when they are integrated effectively into an existing workflow. The relevant unit of analysis is therefore increasingly the AI-enabled process, rather than the individual AI task.

This reinforces the importance of organisational preparedness identified in the banking literature. Moharrak and Mogaji (2025) demonstrate that the adoption of generative AI is influenced by factors including reliability, regulatory requirements, data privacy and organisational responsiveness. Vuković, Dekpo-Adza and Matović (2025) similarly identify governance, explainability, regulation and organisational considerations as important themes in the adoption of AI across financial services. The Raiffeisen case can therefore be understood as an example of a broader principle: technological capability creates potential, but organisational capability determines whether that potential can be converted into sustainable value.

The quality and governance of organisational data are particularly important in this respect. Generative AI systems are only as useful as the information environment within which they operate. Enterprise applications such as Copilot can potentially improve access to internal knowledge, but this benefit depends on information being accurate, current, appropriately structured and accessible under clearly defined permissions. AI adoption can therefore expose weaknesses that already exist within an organisation's information architecture. Poorly governed data may not simply reduce the usefulness of AI; it can cause AI systems to reproduce or amplify existing organisational weaknesses.

The same principle applies to the relationship between AI and human expertise. The evidence discussed in Chapter 6 suggests that human–AI collaboration can produce significant productivity gains, but that the effects vary across employees and tasks. AI can support less experienced workers particularly effectively, while experienced professionals may derive different benefits and may sometimes be exposed to risks associated with inappropriate reliance on AI (Brynjolfsson, Li and Raymond, 2025). This suggests that the competitive advantage created by AI will not depend exclusively on access to advanced models. It will also depend on whether organisations can develop employees who understand when AI should be used, when its outputs should be challenged and when human judgement must take precedence.

The concept of calibrated trust is therefore central to the discussion. Effective human–AI collaboration does not require employees either to trust or distrust AI categorically. Rather, trust should correspond to the demonstrated capabilities and limitations of the system in a particular context. The findings of Dell'Acqua et al. (2025) are particularly relevant here because they show that the same AI system can improve performance on some tasks while reducing it on others. This implies that effective AI adoption requires employees to recognise the boundaries of AI competence rather than assuming that a generally capable system is universally reliable.

This issue becomes increasingly important as AI moves towards greater autonomy. Generative AI used as an assistant leaves the employee largely responsible for initiating and evaluating the task. Agentic AI potentially changes this relationship by allowing systems to perform multiple steps and initiate actions. As discussed by Ganesh et al. (2025), this creates the possibility of substantial augmentation of financial knowledge work, but it also introduces a need to determine appropriate levels of autonomy and human involvement. The transition towards agentic AI should therefore not be interpreted simply as a technological progression in which more autonomy is always better. Rather, autonomy should be matched to the risk and consequences of the activity being performed.

This creates what can be described as an autonomy–control trade-off. Greater autonomy may reduce coordination costs, accelerate processes and increase scalability. However, it also increases the potential consequences of errors and makes effective oversight more difficult. In low-risk and highly structured activities, substantial automation may be appropriate. In activities involving sensitive customer information, regulatory interpretation or significant financial consequences, greater levels of human review may remain necessary. The appropriate degree of autonomy is therefore context-dependent rather than technologically predetermined.

The risks discussed in Chapter 5 reinforce this conclusion. Generative AI introduces concerns surrounding hallucinations, data privacy, cybersecurity, bias, regulatory compliance and accountability. These risks become potentially more consequential when AI systems move from producing information to taking action. An inaccurate AI-generated summary can be corrected before being used; an autonomous agent that acts upon incorrect information may propagate an error through several stages of a process. Agentic AI therefore changes not only the potential benefits of AI but also the magnitude and nature of its risks.

The regulatory environment further constrains the speed and extent of this transition. Banking is characterised by requirements concerning operational resilience, data protection, consumer protection, model risk and accountability. Consequently, AI systems cannot be evaluated solely according to their technical performance. Their deployment must also be assessed according to whether their behaviour can be monitored, their outputs validated and their actions attributed to accountable individuals or organisational processes. Crisanto et al. (2024) emphasise the importance of governance, data management, model risk and third-party dependencies in the financial-sector adoption of AI. These concerns become increasingly relevant as financial institutions move towards more autonomous AI systems.

An important implication is that AI governance should not be viewed as a constraint external to innovation. Effective governance can instead be understood as an enabling capability. Clear data permissions, defined accountability, appropriate human approval mechanisms and reliable monitoring can create the conditions under which organisations are able to deploy AI more confidently. The objective is therefore not to eliminate risk, which would make meaningful innovation impossible, but to ensure that the level of AI autonomy remains proportionate to the risks associated with a particular process.

This perspective also changes how the Raiffeisen case should be evaluated. The strategic significance of its adoption of generative AI, Copilot and agentic AI does not lie simply in being an early adopter of emerging technologies. The more important question is whether Raiffeisen can develop the complementary organisational capabilities required to use these technologies effectively. These include AI literacy among employees, robust data governance, workflow redesign, appropriate risk management, mechanisms for human oversight and a culture capable of learning from AI deployment.

The case also illustrates why AI transformation should not be understood as a linear technological process. Moving from generative AI to agentic AI does not necessarily represent progress if the organisation has not developed the governance and organisational capabilities required to manage greater autonomy. An institution may obtain more value from a well-governed AI assistant embedded in a high-value workflow than from an autonomous agent deployed in a poorly understood or inadequately controlled process. The maturity of the organisation may therefore be as important as the maturity of the technology.

The broader literature supports this interpretation. Eisfeldt and Schubert (2025) identify finance as an industry with substantial exposure to the productivity effects of generative AI because of its high concentration of information-intensive work. Ganesh et al. (2025) extend this argument by identifying the potential of AI agents to augment financial knowledge work. Aldasoro et al. (2025) further demonstrate that AI agents may perform sophisticated financial tasks in controlled environments. Taken together, these studies suggest that the technological potential is significant. However, they do not imply that financial institutions should maximise AI autonomy irrespective of context. Rather, the emerging evidence points towards a future in which value will depend on carefully designed combinations of AI capability and human expertise.

The Raiffeisen case can consequently be interpreted as an early example of this broader organisational transition. Generative AI and Copilot provide mechanisms for augmenting employees and improving access to information, while agentic AI introduces the possibility of extending AI participation from information production towards process execution. The case therefore illustrates both sides of the transformation: AI can become increasingly capable of performing work, while organisations must simultaneously become increasingly capable of supervising, integrating and governing AI.

The central finding of this discussion is therefore that the strategic challenge facing banks is not simply one of technological adoption. It is one of organisational adaptation. Banks that fail to adopt useful AI capabilities may sacrifice potential productivity and innovation gains. However, banks that pursue autonomy without sufficiently developed governance, data infrastructure and human capabilities may create new operational, regulatory and reputational risks. The relevant competitive advantage is consequently unlikely to come from AI adoption alone, because access to increasingly powerful foundation models will become progressively more widespread. Instead, advantage is more likely to arise from the ability to integrate AI effectively into proprietary processes, organisational knowledge and human expertise.

For Raiffeisen, this implies that the transition towards agentic AI should be approached as a process of controlled organisational experimentation rather than as a simple technology rollout. Use cases should be evaluated according to their potential value, risk, data requirements and appropriate level of human involvement. Employees should be trained not merely to operate AI systems but to evaluate their outputs and recognise their limitations. At the same time, governance mechanisms should evolve alongside technological capability so that increasing AI autonomy is accompanied by corresponding increases in monitoring, accountability and control.

The research question posed in the introduction can therefore be answered in two parts. Generative AI and agentic AI have the potential to transform banking by augmenting knowledge work, improving information access, reducing coordination costs and increasingly automating multi-step processes. However, the realisation of these benefits depends on organisational preparedness, data quality, employee capabilities, calibrated human oversight and effective governance. The central challenge for institutions such as Raiffeisen is consequently not to choose between innovation and control, but to develop an organisational environment in which innovation is made sustainable through control.

The transformation of banking through AI should therefore be understood neither as a straightforward path towards automation nor as a simple extension of existing digitalisation. It represents a reconfiguration of the relationship between humans, information and organisational action. The long-term significance of Raiffeisen's AI strategy will ultimately depend on whether it can translate increasingly capable technologies into reliable, governed and genuinely value-creating forms of human–AI collaboration.

8. Conclusion

Artificial intelligence is becoming an increasingly important component of the transformation of financial services. The analysis presented in this paper demonstrates that the significance of current AI developments extends beyond the automation of individual banking tasks. Generative AI is beginning to reshape knowledge work by supporting information retrieval, summarisation, analysis, communication and decision preparation, while agentic AI introduces the possibility that AI systems will increasingly perform and coordinate sequences of actions within organisational workflows. The Raiffeisen case provides a relevant illustration of this transition from AI as a productivity tool towards AI as an increasingly integrated component of organisational processes.

The analysis nevertheless indicates that technological capability alone does not determine the value of AI. Recent empirical research shows that the effects of generative AI vary according to the nature of the task, the experience of the employee and the relationship between human and AI capabilities (Brynjolfsson, Li and Raymond, 2025; Dell'Acqua et al., 2025). AI can generate substantial productivity improvements, but it can also reduce performance when used outside its capability frontier or when employees rely on it without sufficient critical evaluation. Consequently, the relevant question for financial institutions is not simply whether AI can perform a task, but whether its use improves the overall process while maintaining appropriate standards of reliability, accountability and control.

The potential of agentic AI makes this distinction even more important. By combining information retrieval, reasoning, tool use and action, AI agents could potentially transform how complex financial processes are organised. Applications may extend from employee assistance and compliance support to financial analysis, operational processes and liquidity management (Ganesh et al., 2025; Aldasoro et al., 2025). However, increased autonomy also increases the consequences of failure. An incorrect generated response can be reviewed before use; an autonomous system capable of executing multiple actions may instead propagate an error across an entire workflow. The transition towards agentic AI therefore represents not simply an increase in technological capability but an increase in the importance of governance.

The risks identified in this paper demonstrate why responsible implementation is essential. Hallucinations and unreliable outputs can undermine decision quality, while extensive access to organisational data creates privacy and confidentiality concerns. Generative AI can strengthen cybersecurity capabilities but may simultaneously create new vulnerabilities, including more sophisticated social-engineering attacks (Aldasoro et al., 2024). Regulatory and accountability challenges become particularly significant when AI systems influence consequential decisions or interact with external systems. Furthermore, emerging evidence suggests that widespread adoption of generative AI may have implications beyond individual institutions, including potential effects on systemic financial risk (Zhao, Dai and Nagayasu, 2025). AI governance must therefore address not only model performance but also data, human behaviour, organisational processes and systemic dependencies.

The analysis also demonstrates that human–AI collaboration should not be reduced to the principle of keeping a human “in the loop”. Effective human oversight requires employees to possess the knowledge and authority necessary to evaluate and challenge AI outputs. The evidence on the uneven effects of AI across tasks demonstrates that employees need to understand both the capabilities and limitations of the technology (Dell'Acqua et al., 2025). At the same time, organisations must consider how AI affects employee motivation, professional development and the acquisition of expertise. AI may reduce routine cognitive work, but extensive automation of foundational tasks could also alter how employees develop professional competence. Successful implementation therefore requires investment in AI literacy, organisational learning and redesigned roles rather than simply providing employees with access to AI tools.

For Raiffeisen, these findings imply that the strategic challenge is to develop AI capabilities and governance capabilities simultaneously. The adoption of generative AI and Copilot can provide an important foundation by increasing employee productivity and improving access to organisational knowledge. The subsequent development of agentic AI should, however, be approached through controlled and risk-sensitive experimentation. Low-risk, structured activities may be suitable for relatively high levels of automation, whereas processes involving sensitive data, material financial consequences, regulatory interpretation or significant customer impact are likely to require stronger human oversight. The appropriate level of AI autonomy should therefore be determined by the characteristics and consequences of the task rather than by the technological capabilities of the system alone.

This leads to the central conclusion of the paper: AI transformation in banking is fundamentally an organisational challenge rather than merely a technological one. As increasingly capable AI becomes more widely available, access to the underlying technology is unlikely to remain a decisive source of competitive advantage. Instead, differentiation is likely to depend on how effectively financial institutions combine AI capabilities with proprietary data, institutional knowledge, employee expertise, workflow design and governance. Banks that successfully develop these complementary capabilities may be able to achieve significant productivity and innovation benefits while maintaining appropriate levels of control.

The Raiffeisen case therefore illustrates an important transition in the future of banking. The question is no longer whether financial institutions will use AI, but how deeply AI will become embedded in their organisational processes and how much autonomy institutions will be willing to delegate to AI systems. The answer should not be unlimited automation. Rather, the evidence supports a model of risk-proportionate autonomy, in which the degree of AI independence increases where processes are sufficiently structured, controllable and low-risk, while human judgement and accountability remain central where consequences are significant.

Ultimately, the future of AI in banking will depend on the ability to reconcile two objectives that can appear to be in tension: innovation and control. Excessive caution may prevent institutions from capturing significant productivity and innovation opportunities, while uncontrolled adoption may create operational, regulatory and reputational vulnerabilities. The appropriate objective is therefore not to choose between these alternatives but to develop the organisational capabilities that allow them to reinforce one another. Sustainable AI transformation occurs when increasing technological capability is matched by increasing organisational capacity to govern, supervise and work with it.

References

Aldasoro, I., Doerr, S., Gambacorta, L., Notra, S., Oliviero, T. and Whyte, D. (2024) ‘Generative artificial intelligence and cyber security in central banking’, BIS Papers, No. 145, Bank for International Settlements.

Aldasoro, I. and Desai, A. (2025) ‘AI agents for cash management in payment systems’, BIS Working Papers, No. 1310.

Aldasoro, I. et al. (2024) ‘Generative artificial intelligence and cyber security in central banking’, Journal of Financial Regulation, 11(1), pp. 119–143.

Brynjolfsson, E., Li, D. and Raymond, L.R. (2025) ‘Generative AI at Work’, The Quarterly Journal of Economics, 140(2), pp. 889–942. doi: 10.1093/qje/qjae044

Crisanto, J.C., Leuterio, C.B., Prenio, J. and Yong, J. (2024) ‘Regulating AI in the financial sector: Recent developments and main challenges’, FSI Insights, No. 63. Bank for International Settlements.

Dell'Acqua, F., McFowland, E. III, Mollick, E., Lifshitz, H., Kellogg, K.C., Rajendran, S., Krayer, L., Candelon, F. and Lakhani, K.R. (2026) ‘Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality’, Organization Science, 37(2), pp. 403–423. doi: 10.1287/orsc.2025.21838.

Eisfeldt, A.L. and Schubert, G. (2025), “Generative AI and Finance”, Annual Review of Financial Economics, 17, pp. 363–393.

Ganesh, S. et al. (2025) ‘Generative AI agents for knowledge work augmentation in finance’, Annual Review of Control, Robotics, and Autonomous Systems, 8, pp. 189–210.

Baffour Gyau, E., Appiah, M., Gyamfi, B.A., Achie, T. and Naeem, M.A. (2024), ‘Transforming banking: Examining the role of AI technology innovation in boosting banks financial performance’, International Review of Financial Analysis, 96, 103700. DOI: 10.1016/j.irfa.2024.103700. It is a peer-reviewed journal article.

Dietrich, Andreas ‘KI bei Raiffeisen: So nutzt die Bankengruppe Generative AI, Copilot und Agentic AI’. Lucerne University of Applied Sciences and Arts.

Moharrak, M. and Mogaji, E. (2025) ‘Generative AI in banking: Empirical insights on integration, challenges and opportunities in a regulated industry’, International Journal of Bank Marketing, 43(4), pp. 871–896.

Singh, R.K., Mishra, R., Kumar, S. and Bag, S. (2025), “Applications of Artificial Intelligence for Optimizing Banking and Financial Operations: A Systematic Literature Review and Future Research Agenda”, Journal of Economic Surveys, 39(5), pp. 2284–2302.

Vuković, D.B., Dekpo-Adza, S. and Matović, S. (2025) ‘AI integration in financial services: A systematic review of trends and regulatory challenges’, Humanities and Social Sciences Communications, 12, Article 562.

Zhao, Y., Dai, R. and Nagayasu, J. (2025), ‘Generative AI: The transformative impact of ChatGPT on systemic financial risk in Chinese banks’, Pacific-Basin Finance Journal, 93, 102829.

Contact

Reach out via email for inquiries.

Email

Subscribe to newsletter

info@grcadvisory.ch

© 2025. All rights reserved.